South Korea Warns Against AI Mental Health Diagnoses as Survey Finds Growing Dependence

One in Three Users Reporting Poor Mental Health Accepted AI Diagnoses

South Korea has issued new guidance against treating chatbot conversations as medical diagnoses after a government survey found that 33.9% of respondents who rated their mental health as poor had diagnosed themselves with a mental illness using AI or accepted its answers as medical judgments. Among all respondents, the proportion was 14.6%, while just 11.9% had checked AI mental health advice with a professional or official information.

The Ministry of Health and Welfare and the Korean Neuropsychiatric Association released the mental health AI guidelines on Thursday, October 8, 2026, ahead of World Mental Health Day on October 10. They contain 21 recommendations for general users, plus seven each for children and adolescents and their guardians.

A further finding gives the warning particular urgency: among 115 respondents who had disclosed psychological distress, confusion or serious situations involving depression, self harm or suicidal impulses to AI, 49.6% took no separate action beyond ending the conversation. The guidance directs people facing a mental health crisis toward trusted people and professional services rather than continued reliance on a chatbot.

The policy does not ban emotional conversations with AI, nor does it certify particular services as safe. It draws a boundary between obtaining information or comfort and handing a machine responsibility for diagnosis, treatment or major personal decisions.

Lee Sun-young, director-general for mental health policy at the ministry, explained why a reassuring tone should not be mistaken for medical reliability:

Generative AI can serve as a new tool for obtaining information and emotional support, but friendly and confident responses do not necessarily guarantee accuracy or safety.

What the Survey Measures

The Korea Institute for Health and Social Affairs supervised the online survey, which Korea Research conducted from August 25 to September 4, 2026. It covered 1,019 people aged 15 to 79 who had used generative AI within the previous month.

That sampling boundary matters. These findings describe recent AI users, not the entire South Korean population. The respondents also reported their own experiences and perceptions. The survey does not establish that AI caused mental illness or that people who described themselves as dependent had a clinically diagnosed condition.

Within the sample, 112 respondents rated their mental health as poor. Their 33.9% rate of accepting AI diagnoses compared with 18.2% among those who rated their mental health as average and 9.3% among those who rated it as good. The poor mental health group therefore reported this behavior at about 3.6 times the rate of the good mental health group.

The direction of that relationship cannot be settled by the figures alone. People already struggling may turn more readily to chatbots, while chatbot interactions may affect how they understand their symptoms. What the survey establishes is that the users reporting greater distress were also more likely to treat AI responses as medical conclusions.

Why People Confide in Chatbots

Most respondents, 81.3%, first encountered generative AI through study, work, information searches or acquiring knowledge. Its use had since extended into emotional life: 68.8%, or 701 people, said they used it for everyday conversation, emotional connection, comfort or psychological counseling.

The current purposes overlapped. Some 54.4% used AI for daily conversation or companionship, 48.4% for emotional connection and comfort, and 44.1% for mental health or psychological counseling. These percentages should not be added together because respondents could select more than one purpose.

Among emotional users, 85.6% valued being able to talk anytime and anywhere, and 77.6% cited the ease of ending a conversation. Another 71.9% wanted to avoid gossip, 70.9% were reluctant to discuss the subject with people they knew, and 62.6% did not want to worry others.

Many described practical benefits. Across all respondents, 79.9% said AI answers or advice had helped them or resolved a problem. More specifically, 54.9% said conversations helped them organize their situation, 49.7% said they helped them reflect on the causes of emotions, and 49.8% learned about mental health information or services.

Those are reports of perceived usefulness, not clinical measurements of recovery. Still, they help explain why the government chose guidance on safer use rather than a blanket rejection of the technology.

When Advice Becomes a Diagnosis

The ministry warns against using AI conversations to conclude that someone has depression, an anxiety disorder or another mental illness. A chatbot's assessment of symptoms or risk should not be accepted as a definitive medical judgment.

The survey found that 16.5% of respondents had mistaken confident AI responses, including false or harmful advice, for expert opinions. That result points to a problem distinct from factual accuracy: an answer can sound authoritative even when its reasoning is unreliable or its advice does not fit the person's circumstances.

The guidelines recommend checking mental health information against verified material from medical societies and government agencies, or asking qualified professionals. They also advise users to consider other explanations and perspectives rather than treating a chatbot's agreement as proof that their own interpretation is correct.

Choi Joon-ho, president of the Korean Neuropsychiatric Association, described the scope of a proper assessment:

Mental health problems should be assessed by considering an individual's life circumstances, interpersonal relationships, changes in symptoms and their impact on daily functioning, rather than focusing on a single symptom.

This distinction matters because a conversation about one symptom may leave out its duration, changes over time, other health conditions and effects on daily life. Choi also stated:

AI cannot replace professional evaluation and treatment, making it important for people to seek timely medical assistance when necessary.

Dependence Can Affect Sleep and Major Decisions

Among all respondents, 16.3% said they continued talking to AI even when they wanted to stop and focus on activities outside the conversation. Another 9.8% reported that difficulty controlling conversation time had disrupted sleep or daily life.

These experiences were more common among the 449 respondents who used AI specifically for mental health or psychological counseling. In that group, 23.4% had difficulty stopping conversations and 15.1% reported disruptions to sleep or daily activities.

The difference also extended to major decisions. Across the full sample, 10.6% had acted on AI advice about matters such as resigning from a job or ending a relationship without consulting other people. Among counseling users, the proportion was 17.8%, a difference of 7.2 percentage points.

The recommendations therefore go beyond limiting screen time. Users should decide beforehand what they will discuss and when they will stop, avoid treating AI as a real person, and reduce interactions if chatbot conversations begin displacing relationships with family or friends. Important decisions should be discussed with trusted people or qualified professionals.

The crisis findings show why ending a conversation may not be enough. Among 146 respondents who experienced unpleasant or harmful AI answers, 45.9% brushed them aside or continued talking. The guidance instead recommends stopping dangerous interactions and seeking human help. It identifies South Korea's suicide prevention line, 109, and mental health crisis counseling line, 1577-0199, as assistance routes.

Perceived Privacy Is Not Data Protection

The attraction of talking without fear of gossip does not mean that a conversation is confidential. The ministry found that 38.0% of respondents worried their discussions could be leaked or used for other purposes.

Protective behavior was uneven. Some 61.0% filtered out private information they would not tell other people, 56.6% anonymized or omitted details identifying a person or organization, and 50.8% changed the subject or stopped when they felt they were sharing too much.

The guidance advises checking a service's purpose, intended audience, evidence of effectiveness and personal data practices before using it for mental health support. Where a service claims medical benefits, users should check whether recognized authorities have verified or approved it.

Names, telephone numbers and addresses should not be entered, nor should sensitive documents such as diagnostic certificates, prescriptions or test results. For young users, the recommendations also identify school names and photographs as information to protect.

Data protection was a policy priority for respondents too: 66.8% wanted rules governing the storage and use of conversation records and guarantees of user control. The connection is direct. Advice to protect personal information is more useful when services also explain what they collect, retain and reuse.

Teenagers Need More Than Guardian Awareness

Among the 201 respondents aged 15 to 18, 69.2% used AI for emotional purposes, close to the 68.8% rate across the full sample. The more striking result was dependence: 51.7% described themselves as dependent on AI, the highest share of any age group.

The ministry describes that result as perceived dependence. Some reports describe the same figure more specifically as relying on AI more than on other people. Neither description should be treated as a clinical diagnosis of addiction.

Although 83.1% of adolescent respondents said their guardians knew they used AI, only 2.5% had agreed usage rules with a guardian. That is an 80.6 percentage point gap between awareness and shared rules, though the figures do not show that awareness alone prevents harmful use.

The separate youth recommendations ask children to consider their own thoughts and circumstances before following AI advice, recognize that they can stop a conversation whenever they choose, and check whether they are treating AI characters as real friends or romantic partners. Dangerous responses should be shared with trusted adults rather than handled alone.

Guardians are advised to listen to how and why children use AI, check age suitability and privacy settings, and agree rules together instead of imposing blanket bans. They should watch for growing use, shrinking human relationships and signs of crisis. School counselors, Wee Class counseling centers and youth counseling services, including 1388, are identified as support options.

The warning arrives amid a broader rise in youth treatment: psychiatric care for people under 18 increased by 76.6% between 2020 and 2024. Separate findings on student sleep deprivation showed 46.7% of students at general high schools sleeping fewer than six hours. Those studies do not establish a connection to chatbot use, but they explain why protecting sleep and access to professional care matters in the adolescent guidance.

Clinical Evidence Does Not Validate Every Chatbot

A digital psychiatry review published online in Clinical Psychopharmacology and Neuroscience on October 1, 2025, describes both potential benefits and serious limitations. It examines trials and other research on chatbots designed for mental health interventions, rather than treating every conversational system as equivalent.

The review cites a randomized trial of Therabot reporting average symptom reductions of 51% for major depressive disorder and 31% for generalized anxiety disorder. It also discusses structured tools drawing on cognitive behavioral therapy, which helps people examine links between thoughts, feelings and behavior.

Those findings cannot establish that a general chatbot provides reliable diagnoses or safe treatment. Products differ in design, therapeutic content and supervision. The review itself identifies limited evidence on lasting effects and the need for stronger testing of generative systems for false information, inconsistent responses and unsafe behavior.

Its safety concerns include harmful responses during crises, reinforcement of delusional beliefs, bias and privacy risks. It recommends crisis procedures supervised by people, stronger governance of mental health data and care that combines AI tools with clinicians rather than replacing them.

What Government Action Comes Next

The ministry's announcement follows earlier policy commitments and sets a timetable for more targeted guidance:

• January 2025: South Korea issued Generative AI Medical Device Approval and Review Guidelines, discussed in the academic review.
• March 27, 2026: The Third Basic Plan for Mental Health Welfare, covering 2026 to 2030, included plans for AI dependence guidance.
• June 9, 2026: Government measures on teenage suicide prevention also set out plans for guidance.
• August 25 to September 4, 2026: The survey of 1,019 recent generative AI users was conducted.
• October 8, 2026: The ministry and psychiatric association released the user and guardian recommendations.
• 2027: The ministry plans tailored guidance for groups at greater risk, including people with severe mental illness.

Public support for intervention was broad: 86.9% said government had a role in addressing problems associated with AI use. That included 45.0% favoring guidance and 41.9% favoring legal regulation. Some 68.7% identified mental health AI guidelines and changes to legal frameworks as urgent priorities.

User guidance and product oversight address different responsibilities. The World Health Organization's January 2024 guidance on large multimodal models, systems that process information such as text and images, warned about automation bias, meaning excessive trust in machine output. It recommended independent audits and impact assessments after large scale deployment, with results made public.

South Korea's new recommendations are not an evaluation or endorsement of individual services. The ministry plans educational programs for schools and clinical settings, but the announcement gives no exact release date for those programs or the 2027 guidance. It also does not announce a new binding law governing all emotional support chatbots.

Key Points

• South Korea released mental health AI guidance on October 8, 2026, with 21 general recommendations and seven each for young users and guardians.
• Among respondents reporting poor mental health, 33.9% accepted AI assessments as diagnoses or used them to diagnose themselves.
• Only 11.9% of all respondents checked AI mental health advice with professionals or official information.
• Difficulty stopping conversations and acting solely on AI advice were more common among psychological counseling users.
• Among teenagers surveyed, 51.7% reported dependence, while only 2.5% had agreed usage rules with guardians.
• The guidance calls for privacy protection, human consultation and professional help in crises, with tailored recommendations planned for 2027.